GEOGRAPHIC INFORMATION ISSUES ASSOCIATED WITH SOCIOECONOMIC MODELLING FROM NIGHT-TIME LIGHT REMOTE SENSING DATA
C. N. H. Doll a, *, J. G. Morley a, J-P. Muller a
a
Dept. of Geomatic Engineering, University College London, Gower Street, London WC1E 6BT, United Kingdom.
(cdoll, jmorley, jpmuller) @ge.ucl.ac.uk
Commission VII, WG VII/6
KEY WORDS: Remote Sensing, GIS, Human Settlement, Urban, Mapping, Modeling, Economy
ABSTRACT:
The ability to model socio-economic parameters from remote sensing data sources is an important element in addressing the data
gap that currently exists for linking human and physical systems in Earth system models. The provision of information about such
parameters in a gridded format would be highly advantageous for incorporation with other environmental datasets within a
Geographical Information System (GIS) environment. This paper outlines some of the geographical information issues encountered
in a recent study to analyse and map the relationship between night-time radiance from the Defense Meteorological Satellite
Program’s Operational Linescan System (DMSP-OLS) and Gross Domestic Product (GDP) for countries in the European Union
and the United States. Relationships between total radiance and GDP were constructed at a number of sub-national spatial scales.
These relationships were then used to create maps of economic activity for these regions at an enhanced spatial resolution of 5km.
The use of calibration statistics from administrative areas of different sizes gave rise to issues such as the Modifiable Area Unit
Problem (MAUP) and the Ecological Fallacy. Correlation coefficients of these relationships were found to be highly variable as a
result of the MAUP. Analysis of different economic sectors revealed that the overall relationship is a combination of different
relationships for the different economic sectors present in any given administrative area. This paper not only intends to highlight
how one may use remote sensing data to successfully create a map of economic activity compatible for use within a GIS, but also to
demonstrate the complementarity between remote sensing and GIS at every level of analysis.
1. INTRODUCTION
Night-time light imagery can be intuitively perceived to be
indicative of a range of socio-economic parameters. Socioeconomic data are collected at irregular spatial units. Nighttime light imagery potentially offers a means to disaggregate
these data to a constant spatial resolution. Parameters such as
human population, Gross Domestic Product (GDP), power
consumption and even greenhouse gas emissions all have
strong correlations with lit-area and/or total radiance (Elvidge
et al., 1997; 1999; 2000; Doll et al., 2000). Previous work used
country level aggregations of light and GDP to build a global
relationship. This paper investigates the characteristics of subnational relationships and discusses the implications for
producing disaggregated maps of these parameters. The main
body of results pertaining to both the relationships and maps
are presented in an expanded paper submitted to Ecological
Economics (Doll et al., 2004).
1.1 Issues surrounding multi-scale analysis of night-time
light data.
The use of multi-scale data introduces issues, which ought to
be taken into consideration when results are analysed. In
particular the Modifiable Areal Unit Problem (MAUP)
(Openshaw, 1984) and the ecological fallacy (Cao and Lam,
1997) are two complementary themes which frequently occur
in GIS analysis, but are usually overlooked due to the lack of
general robust solution. The MAUP consists of two separate
* Corresponding author.
but closely related components, which have been observed to
cause a variation in statistical results depending on the spatial
arrangement of zones or the scale used to analyse the data.
These are known as scale and zoning effects. The scale effect
occurs when different results are observed from the same data
at different spatial resolutions (Wrigley et al., 1996). This
arises due to the aggregation of data into larger units (e.g.
enumeration districts – wards – counties – regions). In addition
to this, different results can also be produced where the scale
of analysis remains constant but constituent areas are
aggregated in a different way or zone boundaries are altered.
These two issues are particularly pertinent in this analysis
since the relationships derived from the data will be used as an
input to the map production procedure.
A practical application of using the zoning effect can be seen
as far back as the early 19th century, where it was observed
that voting districts could be divided in such a way so as to
concentrate the power of the ruling party. This occasionally
gave rise to bizarrely shaped districts epitomised by that of a
former governor of Massachusetts, Elbridge Gerry, whose
party created a salamander shaped district in 1812. The term
gerrymander was coined to describe the elected member from
such a district as well as the action of altering district
boundaries to gain an electoral advantage.
The MAUP has only been intermittently studied. It was
identified by Gehlke and Biehl (1934 cited in Openshaw,
1984) and was again revisited in 1984 by Stan Openshaw.
Whilst its effect has been well documented, it is also an issue
that is frequently overlooked in spatial analysis, since its effect
can take many forms. Openshaw (1984) remarked that ‘the
aggregational variability is not susceptible to a statistical
approach since no systematic regularities could be found’. A
corollary to the MAUP is the ecological fallacy.
In an analytical environment where results are known to vary
depending on the scale or zoning structure, it follows that
erroneous inferences can be made from data that are
generalised from one scale to another. It has three
manifestations depending on the direction of the scale change.
The individualistic fallacy occurs when a macro-level
relationship is imputed from a micro-level (individual)
relationship. Cross-level fallacies describe those inferences
made about one sub-population from another at the same scale
of analysis. The final fallacy, the ecological fallacy is the
opposite of the individualistic fallacy and describes those
inferences made about the fine resolution from the coarse
resolution (Cao and Lam, 1997). In this sense, the cross-level
fallacy arises from the zoning effect, whilst the individual and
ecological fallacies are derived form the scale effect. A multiscale approach is anticipated to facilitate the identification of
MAUP effects that exist with this type of analysis and what
implications this has with respect to creating an accurate map.
2. DATA SOURCES
Satellite data from the Defense Meteorolical Satellite
Programme’s Operational Linescan System (DMSP-OLS) were
used in conjunction with economic and energy consumption
figures for sub-national administrative units.
2.1 DMSP-OLS Radiance Calibrated Data
DMSP-OLS data products are produced by the National
Oceanic and Atmospheric Administration’s National
Geophysical Data Center (NOAA-NGDC) in Boulder,
Colorado. Temporal composite imagery is used to build up a
global map of night-time radiances over the Earth’s surface.
The radiance calibration is facilitated by using a variable gain
setting on the OLS sensor in order to reduce the effects of
saturation from very bright urban centres (Elvidge et al.,
1999). In this respect, the radiance calibrated dataset is
technically superior to the previous city lights dataset. The
dataset refers to imagery captured between March 1996 and
February 1997.
2.2 Ancillary Statistics
Co-temporal GDP and energy consumption data was obtained
for the United States and 11 European Union (EU) countries.
For the US, Gross State Product (GSP) were downloaded from
the US Bureau of Economic Analysis (BEA, 2002) and state
power consumption data from US Energy Information
Administration (EIA, 2003). A more general term for GSP
would be the GRP (Gross Regional Product). It is introduced
here to define the contribution of a sub-national areal unit such
as a generic administrative region or even square pixel areas to
the national GDP total. The term GSP will only be used when
explicitly referring to a state as opposed to aggregations of
states.
European economic data were purchased from Eurostat, the
statistical body of the European Commission. Statistical
reporting in the European Union is done according to the
Nomenclature of Units for Territorial Statistics (NUTS)
system. The NUTS is a five-level hierarchical classification
based on three regional levels and two local levels. Each
member state is divided into a number of NUTS-1 regions
which in turn are divided into a number of NUTS-2 regions
and so on. There are 78 NUTS-1 regions, 210 NUTS-2 and
1093 NUTS-3 units within the 15 EU countries (Eurostat,
2002). In spite of the desire to create a single, coherent
structure of territorial distribution across Europe, the areal
extent of NUTS regions of a given level can differ substantially
between countries.
The digital boundaries of these regions were downloaded from
UNEP’s Global Resource Information Database (GRID) data
collection in Geneva (UNEP, 2003). The digital dataset of the
US state boundaries was obtained from the US Geological
Survey’s National Atlas (USGS, 2003).
3. SUB-NATIONAL CORRELATION
CHARACTERISTICS: THE UNITED STATES
The log-log correlations between socio-economic parameters
and night-time light imagery have been demonstrated to be
strong at the country level (Doll et al., 2000). Studies of these
correlations at the sub-national level are scarce due to the
paucity of disaggregated socio-economic data. A pilot study
examined the variation in the correlation between light and
GDP at two sub-national units for the UK (Doll, 2003). The
strength of the correlation was found to be stronger at the
county level than the regional level. Comparing countries at
the international level also gives a stronger correlation than at
the regional level. This effect is believed to be a manifestation
of the MAUP.
This section describes the MAUP for correlations involving
night-time light imagery and socio-economic parameters. It
will also make inferences about why these variations occur,
and secondly what it can tell us about the overall relationship
between light and the parameter in question. In building an
understanding of the smaller scale issues, it is hoped that a
further insight will be gained into modelling socio-economic
data from night-time lights.
3.1 State and regional correlations
Multi-scale linear relationships between light parameters and
Gross State Product (GSP – a state’s contribution to national
GDP), and Power consumption were tested for the USA. These
analyses were carried out at the state level. For each state, the
total amount of radiance was calculated and plotted against the
corresponding power consumption or GSP figure.
Data are also presented aggregated up to the regional level as
defined by the US Census regions (see Figure 1). Examining
the relationship between energy consumption, GRP and nighttime lights (Table 1) the state level relationship is consistently
better than the regional level one. The intra-state relationship
is better still for 6 out of 9 regions having a linear R2 value
greater than 0.9. However, intra-regional state correlations
reveal that some regions are highly correlated (R2 > 0.9 for the
southern states). For all cases, correlations using the total
(summed) radiance figures perform better than the lit-area
figures.
US Census Divisions
US States
Lit-area
0.499 / 0.328
0.699 / 0.563
Log lit-area
0.641/ 0.373
0.729 / 0.467
Total Radiance
0.652 / 0.49
0.844 / 0.75
Log radiance
0.784 / 0.55
0.900 / 0.79
Two of the aggregation methods produce ‘regions’ which are
composed of non-contiguous states. Examining how different
aggregations affect the correlation of total radiance with the
GRP, it appears that a wide range of results can be obtained
depending on how one assembles the regions.
3.2 Regional aggregations
In addition to comparing the r-squared value at this scale, the
intra-regional correlation was also computed to see if any
discernable patterns were present. The intra-regional
correlation here refers to the average correlation of states
within a given zone. Figure 2 shows firstly that a regional
correlation of different strengths (0.4 –0.95) can be obtained
depending on how regions are arranged. Secondly, that the
intra-regional correlation (i.e. those states which when
summed form one point on the regional level scatter plot)
declines as the regional level correlation increases.
The variation in the correlation of parameters at a given scale
depends on how regions are aggregated. In having data at two
hierarchical sub-national scales, it is possible to test the
national correlation at the smaller scale (regions), by using
different aggregations of larger scale units (states). Economic
data from the BEA were also presented using their regional
classification. Their grouping of states is similar to that of the
US Census using traditional aggregations based on states’
locations with respect to the various geographical regions (e.g.
Plains, Great Lakes, Mountain Division, South Atlantic). In
addition to these two classical aggregations, three others were
devised. One was a geographical aggregation, which
established five regions according to their latitudinal position.
The other two consisted of a random aggregation based on
seven alphabetical divisions and one, which divided the states
after they had been ranked by GSP. The five aggregations are
shown in (Figure 1).
Figure 2. Comparison of regional and (mean) intra-regional
correlation coefficients (of total radiance vs. GDP by state) for
different aggregation methods.
Table 1. Correlation of energy consumption / GRP and nighttime lights at state and aggregated US Census Division level.
US Census Divisions (9)
Latitudinal Divisions (5)
US BEA (8)
These results suggest that the US Census divisions seem to be
unsuitable for building a national scale correlation, but more
appropriate for intra-regional analysis. By shifting one or two
states here and there to build the BEA regions, the correlations
improve in both measures, though these geographical divisions
are generally unsuitable. Even something as random as an
alphabetical classification provides a better regional
correlation result than the two traditional geographically
regional zones. By using a geographical criterion to aggregate
states, the latitudinal divisions provide a regional and intraregional correlation that is most similar to each other.
However, ranking states based on their GSP gives the best
regional correlation, but the worst intra-regional correlation,
despite the component states being of roughly equal economic
magnitude to each other. The same pattern of results was also
observed for the power consumption data.
Alphabetic zones (7)
4. DERIVED RELATIONSHIPS AND OUTLYING
POINTS
Ranked by Gross State Product (7)
Figure 1. Aggregation of US states into larger zones according
to 5 different criteria.
One further point to examine is what effect different
aggregations have on the magnitude of relationships associated
with these correlations. Figure 3 shows the trajectory of the
relationship for each aggregation method. Also plotted (bold in
red) is the relationship derived from state level radiance-GRP
plot. Two points on the plot are of particular interest. Firstly,
there is a point around 1E+07 radiance units (point A), where
the all-state relationship intersects that of the US Census and
BEA relationships. Secondly, further along around 1.75E+07
radiance units (point B) is a point where all the relationships
converge except the all-state relationship. Despite their
divergence at high radiance values, all models would produce
similar results if the input values were between these points.
The results show that the US Census and BEA relationships
are markedly different from those of the other three
aggregation schemes. More importantly, it also shows that the
overall state-level relationship is most similar to these
simulated aggregation schemes and not to those of the
conventional US regions. This has important implications for
extending this technique to areas that have only one level of
sub-national data.
These values are for relationships constrained to run through
the origin and include the outlying points, whose effect
becomes less influential as the spatial units become larger in
size. The generally good correlation in Figure 4 masks much of
the detail. Despite the standard deviation of the GRP/radiance
quotients being relatively low compared to other countries
analysed the individual NUTS-3 zones have been combined in
such a way as to result in vastly different relationships at
different NUTS levels and confuse the base-level correlation.
Figure 3. Comparison of different relationships derived from
the five aggregation schemes displayed in Figure 1.
Taking these points into consideration it is suggested that the
US has a number of regional sub-economies, hence the high
intra-regional correlation coefficients of the US Census and
BEA divisions. However, the regions themselves vary greatly
and bear little relationship to each other. This is not to say that
there is no general nationwide correlation. The reason why the
intra-regional correlation is so poor for the ranked states may
relate back to the nature of the individual regional economies.
Assembling regions by ranking states puts the most
economically productive states (California, Texas, Florida,
Illinois, Ohio, Pennsylvania, and New York) together.
However, they apparently bear little resemblance to each other.
This trend continues for each group of states. BEA (and US
Census) regions by contrast usually consist of one dominant
state and a number of less economically productive satellite
states. Working down the list of states, each BEA region
generally has a good mix of members from each group of
(ranked) states. The exceptions are the Plains and Mideast
regions which contain Ohio and Illinois, and New York and
Pennsylvania respectively, and conversely the Rocky Mountain
region, which has no member from the most economically
active group of states. However, when viewed in totality their
component states fit a single nationwide model albeit with two
outliers (New York and California). Many countries anlaysed
also exhibited one or two regions which did not fit the general
model for the rest of the country.
Spain provides an interesting example of how an outlying point
can affect its parent regions. The relationships derived at each
NUTS level were found to vary more than any other country.
The gradient becomes progressively shallower as larger areas
are considered in the analysis ranging through y = 0.069x,
0.059x and 0.056x for NUTS levels-3, 2 and 1 respectively.
Figure 4. Night-time lights of Spain with its NUTS-3
boundaries and total radiance – GRP scatterplot. The
relationship is based on NUTS-3 points only, excluding
Barcelona and Madrid. The NUTS-1 regions of Madrid and
Este are outlined in black on the map. Cataluña lies within the
region of Este and is outlined in brown (n=40).
The NUTS-3 area of Barcelona is far more radiant than other
NUTS-3 regions and is observed to be an outlier. The capital,
Madrid is a NUTS-1 region and is about one third brighter
than Barcelona. Barcelona is part of Cataluña, which is itself
part of the Este NUTS-1 region. The other regions of Este have
anomalously low GRP for the total radiance in the region.
Barcelona dominates this region to such an extent that not only
does it pull its NUTS-2 point away from the trendline but it
has also pulled its NUTS-1 point towards it. In fact, it is only
due to the presence of Barcelona in the region that its NUTS-1
point is anywhere near the regression line. Its position without
Barcelona’s contribution is also shown in Figure 4. In this
case, due to the magnitude of these modifiable areal unit
effects, it is most prudent to just use the original NUTS-3
relationship and treat Barcelona and Madrid as separate
outliers. This is the relationship displayed on the graph in
Figure 4. This is the most extreme example of an outlying
point’s effect being carried throughout the aggregation process
encountered for any of the countries analysed and highlights
how points can shift in the plot space according to whether an
economically overactive area is aggregated with points of
average economic activity or below average activity.
to an area. In this example from the Italy, it can be seen that
the services sector has a far higher gradient with total radiance
than for agriculture.
4.1 Characteristics of Outlying points
For a few of the countries tested in the analysis of European
countries, there is a number of points which deviated from the
derived regression line. These points invariably represent
heavily urbanised areas, usually a capital city or its hinterland.
The scatterplot shown in Figure 4 is a visualisation of how the
total amount of radiance in a NUTS region is related to its
GDP output on a national scale. It is a more sophisticated
measure than a simple lit-area versus GDP plot since pixels of
equal areas can have very different radiances. It was hoped that
this feature might help to reduce the magnitude of outlying
regions but it appears that this is still not sufficient for some
countries.
11 EU countries were analysed in a similar manner to that of
Spain. Certain countries exhibit extreme outliers, where
points on the graph lies well above the main trendline. They
are not the most radiant administrative areas but they are the
most economically productive. France and Denmark both had
these types of scatterplots. There is a number of similarities
between the French and Danish outliers. Firstly both Paris and
Copenhagen city centres are themselves NUTS regions. They
are of the order of 100 km2, when the average NUTS-3 region
is 5700km2 in France and 2800 km2 in Denmark. The Ile-deFrance region around Paris is also a NUTS-1 region which is a
fraction of the size of its sister NUTS-1 regions in France. It is
highly urbanised and has a high proportion of higher radiance
pixels. The Ile-de-France region is still an obvious outlier
even after aggregation. However, if it is further combined with
the much larger annular Bassin Parisien region, it begins to
align with the other NUTS-1 regions. If a single scale
independent relationship exists however, then it shouldn’t
matter what spatial units are employed to construct the
relationship.
The difference in the relative size of NUTS regions is an
important issue. Contrasting Spain’s NUTS-3 boundaries with
that of France or Denmark, Spanish areas are reasonably large
and uniform in size – even for Madrid and Barcelona. It begs
the question whether a correlation is only as good as the spatial
units it is based on. Would Barcelona’s NUTS-3 point be an
outlier of the magnitude of Paris or Brussels, if it had a small
administrative boundary in the centre of the city? If so, then
the disaggregation will underestimate the centre of the city and
overestimate the surrounding area by virtue of not having an
appropriate spatial unit to describe its radiance relationship to
economic activity. The same can be said for Milan or Rome,
which also have large NUTS-3 areas when compared to the
concentrated zones of Paris or Copenhagen. Whilst the
problem can be “aggregated away”, the application of a
‘macro-relationship’ to create a disaggregated map would be a
misrepresentation of the finer points of the data.
Examining the relationship between total radiance and
different sectors of the economy (Figure 5), it is clear to see
that different sectors of the make vastly differing contributions
Figure 5. Total Radiance versus GRP by economic sector for
the Italy at NUTS-2 level.
A single point in the total radiance-GRP relationship shown for
Spain in Figure 4 is the combination of each set of points for
the three economic sectors. Most regions have an even mix of
each type of economic sector within a NUTS region, however
very small urban NUTS areas such as Paris or Copenhagen are
exclusively urban and the service sector relationship dominates
the economic make-up of the area generating an outlying point
in the scatterplot.
Since outlying points in the scatterplot are smaller in size as
well as highly urbanised NUTS regions, two additional metrics
may be identified as being potentially helpful to discriminate
these areas. Firstly, the mean radiance of a region could be a
useful feature to discriminate those areas, which, while having
equal total radiance, have different GRP. Secondly, the
proportion of area lit in a NUTS region can give an indication
of how urbanised that region is and therefore an inference can
be made as to which sector of the economy is contributing to
its GDP. The fact that different economic sectors have
different GRP per unit radiance values suggests that a priori
knowledge of these areas (from a land use map for example)
could be used as weighting functions for a region.
5. MAPPING RESULTS FROM RADIANCECALIBRATED DATA
What implications do these observations have for the practical
mapping of economic activity? One key consideration is how
far it is prudent to extend the relationship beyond the known
data range to finer spatial resolutions (The Ecological Fallacy).
Reconciling what the map aimed to display and the technical
specifications of the DMSP-OLS sensor affected the choice of
spatial resolution for the map. Whilst the relationship between
total radiance and GRP has been demonstrated to be consistent
from the NUTS-3 level up to NUTS-1, an unknown factor is
how far the relationship can be extrapolated in the other
direction (NUTS-3 to finer scales) before it becomes invalid.
In addition to this unquantified micro-scale relationship, there
is a desire to maintain an aggregate nature to the GDP figures
in the map. Figure 5 shows that different relationships exist
between the different economic sectors and total night-time
radiance. The overall derived relationship such as that shown
in Figure 4 is an aggregate of these sector combinations.
Keeping the map resolution sufficiently coarse avoids issues of
sector specific GDP representations at finer scales where the
aggregate relationship may not be valid. Although the nighttime data has a nominal resolution of 1km, this is resampled
from 2.7km raw data. This also supports a decision to increase
the spatial resolution of the map by reducing any noise effects
present in the fine resolution data.
Bearing these points in mind, the output resolution was set at
5km as this was reasoned to provide both a detailed map at the
continental scale whilst being coarse enough to generalise
economic activity to a level which shows enough detail in large
cities without compromising small towns. In addition to this is
the advantage of making the mapping process less
computationally intensive. Outlying areas identified for certain
countries were excluded from the main relationship of that
country. However, values were attributed to the radiance cells
in these areas by allocating the remaining value from the
national total to these areas thus ensuring that all of the cells
in a country sum to the published total. The map will over
estimate and under estimate certain areas. This is an inevitable
consequence of the method. Nonetheless it is encouraging to
see that very little adjustment to the value of outlier(s) was
required to constrain the map to a country total. Full details of
the results and maps are presented in Doll et al. (2004)
6. CONCLUSION
Although, the method used to construct the economic activity
map from a given relationship has the inevitable consequence
of over- and underestimation, alternative aggregations of the
finest-scaled zones data did not appear to give substantially
different results compared to the administrative NUTS-1
aggregation from the original data. Estimating country level
GDP in the EU was found to be accurate to within 5% for most
countries and within 7% for all. For a given country-level
percentage error, the method used for the treatment of outliers
will yield an increasing amount of error between its predicted
and observed value as the outlier forms a decreasing
percentage of total GDP for that country.
This is a highly encouraging result when considering that the
map is based on just one variable and re-affirms night-time
light data as a promising tool for mapping socio-economic
parameters. Economic activity mapping could also benefit from
the inclusion of other variables as has been shown with respect
to human population mapping from the Landscan project. The
use of such data would facilitate the construction of ‘poverty
maps’ where GDP/capita can be spatially mapped thereby
highlighting areas of social inequality. An example of this for
Guatemala can be found in Sutton et al. (2004). Indeed, the
application of these datasets are arguably of most use to the
developing world. The importance of other variables such as
land-use and transportation networks is likely to become more
acute as the spatial resolution of the map increases.
Cao, C. and Lam, N, S-N., 1997. Understanding the Scale and
Resolution Effects in Remote Sensing and GIS. In: Quattrochi,
D.A. and Goodchild, M.F. (Eds.) Scale in Remote Sensing and
GIS. CRC Lewis: Boca Raton, FL, USA, 57-72.
Doll, C.N.H., Muller, J-P. and Elvidge, C.D., 2000. Night-time
imagery as a tool for global mapping of socio-economic
parameters and greenhouse gas emissions. Ambio, 29 (3), 157162.
Doll, C.N.H., 2003. Estimating non-population parameters
from night-time satellite imagery. In: Mesev, V. (Ed.)
Remotely Sensed Cities. Taylor & Francis: London, UK, 335354.
Doll, C.N.H., Muller, J-P. and Morley, J.G., 2004. Mapping
regional economic activity from night-time light satellite
imagery. Submitted to Ecological Economics.
EIA, 2003. Energy Information Administration: Multi-state
data.
Available
from:
http://www.eia.doe.gov/emeu/states/multi_states.html
Elvidge, C.D., Baugh, K.E., Kihn, E.A., Kroehl, H.W., Davis,
E.R. and Davis, C.W., 1997. Relation between satellite
observed visible-near infrared emissions, population, economic
activity and electric power consumption. International Journal
of Remote Sensing, 18 (6), 1373-1379.
Elvidge, C.D., Baugh, K.E., Dietz, J.B., Bland, T., Sutton,
P.C. and Kroehl, H.W., 1999. Radiance calibration of DMSPOLS low-light imaging data of human settlements. Remote
Sensing of Enviroment, 68 (1), 77-88.
Elvidge, C.D., Imhoff, M.L. and Sutton, P.C., 2000, Relation
between fossil fuel trace gas emissions and satellite
observations of nocturnal lighting. International Archives of
Photogrammetry and Remote Sensing, GITC bv: Amsterdam,
The Netherlands, 33 (B7), 397-401.
Eurostat, 2002a, Main characteristics of the NUTS. Available
from:
http://europa.eu.int/comm/eurostat/ramon/nuts/mainchar_regio
ns_en.html
Openshaw, S., 1984. The modifiable areal unit problem.
Concepts and Techniques in Modern Geography 38,
GeoBooks: Norwich, UK.
Sutton, P.C., Taylor, M. and Elvidge, C.D. 2004. Urban
populations in the developing and developed countries - the
use of night time data. In: Jurgens, C and Rashed, T (Eds.)
Remote Sensing of Urban and Suburban areas. Kluwer
Acadamical. In Press.
UNEP, 2003a. UNEP-GRID GNV158 - Administrative
Regions/Boundaries
of
Europe.
Available
from:
http://www.grid.unep.ch/data/grid/gnv158.php
USGS, 2003. State Boundaries of the United States.
Available from: http://nationalatlas.gov/statesm.html
REFERENCES
BEA, 2002. Bureau of Economic Analysis: Gross State
Product data. Available from:
http://www.bea.doc.gov/bea/regional/gsp/
Wrigley, N., Holt, T., Steel, D. and Tranmer, M., 1996.
Analysing, modelling and resolving the ecological fallacy. In:
Longley, P. and Batty, M. (Eds.). Spatial Analysis: Modelling
in a GIS Environment. New York, USA: Wiley, 25-41.